Forecasting and Nowcasting

In this chapter

Case Study presentation

Data forecasts by industry sector relating to deflated VAT turnover still unavailable for the “Flash” estimate of Belgian GDP Two-step methodology Pre-selection of a restricted number of predictors (mostly from business surveys) Model selection, definition and prediction Use of the rjdverse packages: rjd3toolkit, rjd3workspace, rjd3providers, rjd3tramoseats, rjd3sts and rjd3nowcasting Types of models considered RegARIMA models: “supervised” TRAMO with or without predictor(s) ADL(1,1), ADL(1,0) with fixed or variable coefficients (predictors) Structural models + regressor(s) with fixed or variable coefficients (predictors) Dynamic factor models (1 factor)

Packages relevant for nowcasting

rjd3toolkit: statistical toolkit and modelling support, for example: calendar_td(): generation of regressors for trading days effects ao_variable() / ls_variable(): generation of regressors for different types of outliers differencing_fast(): automatic identification of differentiation orders to achieve stationarity ljungbox(), jarquebera(), etc.: diagnostics on residuals rjd3workspace & rjd3providers: interactivity with GUI workspaces rjd3tramoseats : RegARIMA models tramo_forecast(): prediction using TRAMO rjd3sts: Structural / State Space models rjd3nowcasting: Dynamic Factor Models

Reg Arimla Models

Forecasting by means of a RegARIMA model (TRAMO) with rjd3tramoseats (full code: here):

GUI

Hybrid GUI/R approach for RegARIMA models (and seasonal [pre-]adjustment)

State Space

State Space modelling: structural models, regression with variable coefficient, etc.

Prediction by means of a Basic Structural Model with regressors: code here

Dynamic Factor models

Conclusion

Cross-validation shows gains in accuracy in almost all branches of activity compared with the previous methodology (“unsupervised” TRAMO, without predictor) Approximatively 2/3 of these gains come from switching to a “supervised” TRAMO method (without predictor) Importance of defining a national calendar for monthly data Importance of properly defining outliers (especially during Covid period) Overall, the predictive power of business surveys turns on to be relatively weak However, this masks disparities and pleasant surprises (e.g. the question on weather conditions in the construction sector)